I've been playing around with K3 a bunch, but the verbosity of the reasoning makes complete e2e agent work basically cost the same as other smaller models, and I'm not seeing a huge difference in quality, just a way longer e2e completion time.

▲

sunaookami a day ago | parent [-]

Same problem with every chinese model currently, they overthink way too much and take too much tokens and time.

	▲	embedding-shape a day ago \| parent \| next [-]
		More or less, yeah. I've found mild success with deepseek-v4-flash though, and also Qwen3.5-122B-A10B-NVFP4 running locally, especially in terms of "doesn't overthink every single prompt" and somewhat reasonable quality. Really wishing for a 3.8 update of the 122B variant, that'd be really competitive (for local usage) :)
	▲	EgregiousCube a day ago \| parent \| prev \| next [-]
		A consequence of aggressive distillation?
	▲	szundi a day ago \| parent \| prev [-]
		[dead]